| blob | 03724992cdf20ea8ed46aaf2aee29ef845f29c54 |
1 /*
2 * raid10.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 2000-2004 Neil Brown
5 *
6 * RAID-10 support for md.
7 *
8 * Base on code in raid1.c. See raid1.c for futher copyright information.
9 *
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/seq_file.h>
30 /*
31 * RAID10 provides a combination of RAID0 and RAID1 functionality.
32 * The layout of data is defined by
33 * chunk_size
34 * raid_disks
35 * near_copies (stored in low byte of layout)
36 * far_copies (stored in second byte of layout)
37 * far_offset (stored in bit 16 of layout )
38 *
39 * The data to be stored is divided into chunks using chunksize.
40 * Each device is divided into far_copies sections.
41 * In each section, chunks are laid out in a style similar to raid0, but
42 * near_copies copies of each chunk is stored (each on a different drive).
43 * The starting device for each section is offset near_copies from the starting
44 * device of the previous section.
45 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
46 * drive.
47 * near_copies and far_copies must be at least one, and their product is at most
48 * raid_disks.
49 *
50 * If far_offset is true, then the far_copies are handled a bit differently.
51 * The copies are still in different stripes, but instead of be very far apart
52 * on disk, there are adjacent stripes.
53 */
55 /*
56 * Number of guaranteed r10bios in case of extreme VM load:
57 */
58 #define NR_RAID10_BIOS 256
66 {
71 /* allocate a r10bio with room for raid_disks entries in the bios array */
77 }
80 {
82 }
84 /* Maximum size of each resync request */
85 #define RESYNC_BLOCK_SIZE (64*1024)
86 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
87 /* amount of memory to reserve for resync requests */
88 #define RESYNC_WINDOW (1024*1024)
89 /* maximum number of concurrent requests, memory permitting */
90 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
92 /*
93 * When performing a resync, we need to read and compare, so
94 * we need as many pages are there are copies.
95 * When performing a recovery, we need 2 bios, one for read,
96 * one for write (we recover only one drive per r10buf)
97 *
98 */
100 {
112 }
116 else
119 /*
120 * Allocate bios.
121 */
127 }
128 /*
129 * Allocate RESYNC_PAGES data pages and attach them
130 * where needed.
131 */
140 }
141 }
145 out_free_pages:
152 out_free_bio:
157 }
160 {
172 }
174 }
175 }
177 }
180 {
188 }
189 }
192 {
195 /*
196 * Wake up any possible resync thread that waits for the device
197 * to go idle.
198 */
203 }
206 {
212 }
215 {
225 /* wake up frozen array... */
229 }
231 /*
232 * raid_end_bio_io() is called when we have finished servicing a mirrored
233 * operation and are ready to return a success/failure code to the buffer
234 * cache layer.
235 */
237 {
243 }
245 /*
246 * Update disk head position estimator based on IRQ completion info.
247 */
249 {
254 }
257 {
266 /*
267 * this branch is our 'one mirror IO has finished' event handler:
268 */
272 /*
273 * Set R10BIO_Uptodate in our master bio, so that
274 * we will return a good error code to the higher
275 * levels even if IO on some other mirrored buffer fails.
276 *
277 * The 'master' represents the composite IO operation to
278 * user-side. So if something waits for IO, then it will
279 * wait for the 'master' bio.
280 */
284 /*
285 * oops, read error:
286 */
293 }
296 }
299 {
310 /*
311 * this branch is our 'one mirror IO has finished' event handler:
312 */
315 /* an I/O failed, we can't clear the bitmap */
318 /*
319 * Set R10BIO_Uptodate in our master bio, so that
320 * we will return a good error code for to the higher
321 * levels even if IO on some other mirrored buffer fails.
322 *
323 * The 'master' represents the composite IO operation to
324 * user-side. So if something waits for IO, then it will
325 * wait for the 'master' bio.
326 */
331 /*
332 *
333 * Let's see if all mirrored write operations have finished
334 * already.
335 */
337 /* clear the bitmap if all writes complete successfully */
344 }
347 }
350 /*
351 * RAID10 layout manager
352 * Aswell as the chunksize and raid_disks count, there are two
353 * parameters: near_copies and far_copies.
354 * near_copies * far_copies must be <= raid_disks.
355 * Normally one of these will be 1.
356 * If both are 1, we get raid0.
357 * If near_copies == raid_disks, we get raid1.
358 *
359 * Chunks are layed out in raid0 style with near_copies copies of the
360 * first chunk, followed by near_copies copies of the next chunk and
361 * so on.
362 * If far_copies > 1, then after 1/far_copies of the array has been assigned
363 * as described above, we start again with a device offset of near_copies.
364 * So we effectively have another copy of the whole array further down all
365 * the drives, but with blocks on different drives.
366 * With this layout, and block is never stored twice on the one device.
367 *
368 * raid10_find_phys finds the sector offset of a given virtual sector
369 * on each device that it is on.
370 *
371 * raid10_find_virt does the reverse mapping, from a device and a
372 * sector offset to a virtual address
373 */
376 {
378 sector_t sector;
379 sector_t chunk;
380 sector_t stripe;
385 /* now calculate first sector/dev */
397 /* and calculate all the others */
403 slot++;
412 slot++;
413 }
414 dev++;
418 }
419 }
421 }
424 {
440 else
442 }
444 }
448 }
450 /**
451 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
452 * @q: request queue
453 * @bvm: properties of new bio
454 * @biovec: the request that could be merged to it.
455 *
456 * Return amount of bytes we can accept at this offset
457 * If near_copies == raid_disk, there are no striping issues,
458 * but in that case, the function isn't called at all.
459 */
463 {
474 else
476 }
478 /*
479 * This routine returns the disk from which the requested read should
480 * be done. There is a per-array 'next expected sequential IO' sector
481 * number - if this matches on the next IO then we use the last disk.
482 * There is also a per-disk 'last know head position' sector that is
483 * maintained from IRQ contexts, both the normal and the resync IO
484 * completion handlers update this position correctly. If there is no
485 * perfect sequential match then we pick the disk whose head is closest.
486 *
487 * If there are 2 mirrors in the same 2 devices, performance degrades
488 * because position is mirror, not device based.
489 *
490 * The rdev for the device selected will have nr_pending incremented.
491 */
493 /*
494 * FIXME: possibly should rethink readbalancing and do it differently
495 * depending on near_copies / far_copies geometry.
496 */
498 {
507 /*
508 * Check if we can balance. We can balance on the whole
509 * device if no resync is going on (recovery is ok), or below
510 * the resync window. We take the first readable disk when
511 * above the resync window.
512 */
515 /* make sure that disk is operational */
522 slot++;
527 }
529 }
531 }
534 /* make sure the disk is operational */
540 slot ++;
544 }
546 }
552 /* Find the disk whose head is closest,
553 * or - for far > 1 - find the closest to partition beginning */
564 /* This optimisation is debatable, and completely destroys
565 * sequential read speed for 'far copies' arrays. So only
566 * keep it for 'near' arrays, and review those later.
567 */
572 }
574 /* for far > 1 always use the lowest address */
577 else
584 }
585 }
587 rb_out:
589 /* conf->next_seq_sect = this_sector + sectors;*/
593 else
598 }
601 {
618 }
619 }
621 }
624 {
629 }
632 {
646 }
647 }
650 }
653 {
654 /* Any writes that have been queued but are awaiting
655 * bitmap updates get flushed here.
656 * We return 1 if any requests were actually submitted.
657 */
667 /* flush any pending bitmap writes to disk
668 * before proceeding w/ I/O */
676 }
681 }
682 /* Barriers....
683 * Sometimes we need to suspend IO while we do something else,
684 * either some resync/recovery, or reconfigure the array.
685 * To do this we raise a 'barrier'.
686 * The 'barrier' is a counter that can be raised multiple times
687 * to count how many activities are happening which preclude
688 * normal IO.
689 * We can only raise the barrier if there is no pending IO.
690 * i.e. if nr_pending == 0.
691 * We choose only to raise the barrier if no-one is waiting for the
692 * barrier to go down. This means that as soon as an IO request
693 * is ready, no other operations which require a barrier will start
694 * until the IO request has had a chance.
695 *
696 * So: regular IO calls 'wait_barrier'. When that returns there
697 * is no backgroup IO happening, It must arrange to call
698 * allow_barrier when it has finished its IO.
699 * backgroup IO calls must call raise_barrier. Once that returns
700 * there is no normal IO happeing. It must arrange to call
701 * lower_barrier when the particular background IO completes.
702 */
705 {
709 /* Wait until no block IO is waiting (unless 'force') */
714 /* block any new IO from starting */
717 /* No wait for all pending IO to complete */
724 }
727 {
733 }
736 {
744 }
747 }
750 {
756 }
759 {
760 /* stop syncio and normal IO and wait for everything to
761 * go quiet.
762 * We increment barrier and nr_waiting, and then
763 * wait until nr_pending match nr_queued+1
764 * This is called in the context of one normal IO request
765 * that has failed. Thus any sync request that might be pending
766 * will be blocked by nr_pending, and we need to wait for
767 * pending IO requests to complete or be queued for re-try.
768 * Thus the number queued (nr_queued) plus this request (1)
769 * must match the number of pending IOs (nr_pending) before
770 * we continue.
771 */
781 }
784 {
785 /* reverse the effect of the freeze */
791 }
794 {
810 }
812 /* If this request crosses a chunk boundary, we need to
813 * split it. This will only happen for 1 PAGE (or less) requests.
814 */
819 /* Sanity check -- queue functions should prevent this happening */
823 /* This is a one page bio that upper layers
824 * refuse to split for us, so we need to split it.
825 */
835 bad_map:
842 }
846 /*
847 * Register the new request and wait if the reconstruction
848 * thread has put up a bar for new requests.
849 * Continue immediately if no resync is active currently.
850 */
863 /*
864 * read balancing logic:
865 */
871 }
887 }
889 /*
890 * WRITE:
891 */
892 /* first select target devices under rcu_lock and
893 * inc refcount on their rdev. Record them by setting
894 * bios[x] to bio
895 */
897 retry_write:
907 }
914 }
915 }
919 /* Have to wait for this device to get unblocked, then retry */
927 }
932 }
955 }
958 /* the array is dead */
962 }
970 /* In case raid10d snuck in to freeze_array */
977 }
980 {
991 else
993 }
1001 }
1004 {
1008 /*
1009 * If it is not operational, then we have already marked it as dead
1010 * else if it is the last working disks, ignore the error, let the
1011 * next level up know.
1012 * else mark the drive as failed
1013 */
1016 /*
1017 * Don't fail the drive, just return an IO error.
1018 * The test should really be more sophisticated than
1019 * "working_disks == 1", but it isn't critical, and
1020 * can wait until we do more sophisticated "is the drive
1021 * really dead" tests...
1022 */
1029 /*
1030 * if recovery is running, make sure it aborts.
1031 */
1033 }
1040 }
1043 {
1051 }
1063 }
1064 }
1067 {
1073 }
1075 /* check if there are enough drives for
1076 * every block to appear on atleast one
1077 */
1079 {
1087 cnt++;
1089 }
1094 }
1097 {
1102 /*
1103 * Find all non-in_sync disks within the RAID10 configuration
1104 * and mark them in_sync
1105 */
1115 }
1116 }
1120 }
1124 {
1133 /* only hot-add to in-sync arrays, as recovery is
1134 * very different from resync
1135 */
1147 else
1154 /* as we don't honour merge_bvec_fn, we must
1155 * never risk violating it, so limit
1156 * ->max_segments to one lying with a single
1157 * page, as a one page request is never in
1158 * violation.
1159 */
1164 }
1173 }
1178 }
1181 {
1194 }
1195 /* Only remove faulty devices in recovery
1196 * is not possible.
1197 */
1202 }
1206 /* lost the race, try later */
1210 }
1212 }
1213 abort:
1217 }
1221 {
1241 }
1243 /* for reconstruct, we always reschedule after a read.
1244 * for resync, only after all reads
1245 */
1249 /* we have read all the blocks,
1250 * do the comparison in process context in raid10d
1251 */
1253 }
1254 }
1257 {
1277 /* the primary of several recovery bios */
1286 }
1287 }
1288 }
1290 /*
1291 * Note: sync and recover and handled very differently for raid10
1292 * This code is for resync.
1293 * For resync, we read through virtual addresses and read all blocks.
1294 * If there is any error, we schedule a write. The lowest numbered
1295 * drive is authoritative.
1296 * However requests come for physical address, so we need to map.
1297 * For every physical address there are raid_disks/copies virtual addresses,
1298 * which is always are least one, but is not necessarly an integer.
1299 * This means that a physical address can span multiple chunks, so we may
1300 * have to submit multiple io requests for a single sync request.
1301 */
1302 /*
1303 * We check if all blocks are in-sync and only write to blocks that
1304 * aren't in sync
1305 */
1307 {
1314 /* find the first device with a block */
1325 /* now find blocks with errors */
1337 /* We know that the bi_io_vec layout is the same for
1338 * both 'first' and 'i', so we just compare them.
1339 * All vec entries are PAGE_SIZE;
1340 */
1344 PAGE_SIZE))
1349 }
1351 /* Don't fix anything. */
1353 /* Ok, we need to write this bio
1354 * First we need to fixup bv_offset, bv_len and
1355 * bi_vecs, as the read request might have corrupted these
1356 */
1374 PAGE_SIZE);
1375 }
1386 }
1388 done:
1392 }
1393 }
1395 /*
1396 * Now for the recovery code.
1397 * Recovery happens across physical sectors.
1398 * We recover all non-is_sync drives by finding the virtual address of
1399 * each, and then choose a working drive that also has that virt address.
1400 * There is a separate r10_bio for each non-in_sync drive.
1401 * Only the first two slots are in use. The first for reading,
1402 * The second for writing.
1403 *
1404 */
1407 {
1413 /* move the pages across to the second bio
1414 * and submit the write request
1415 */
1422 }
1429 else
1431 }
1434 /*
1435 * Used by fix_read_error() to decay the per rdev read_errors.
1436 * We halve the read error count for every hour that has elapsed
1437 * since the last recorded read error.
1438 *
1439 */
1441 {
1450 /* first time we've seen a read error */
1453 }
1460 /*
1461 * if hours_since_last is > the number of bits in read_errors
1462 * just set read errors to 0. We do this to avoid
1463 * overflowing the shift of read_errors by hours_since_last.
1464 */
1467 else
1469 }
1471 /*
1472 * This is a kernel thread which:
1473 *
1474 * 1. Retries failed read operations on working mirrors.
1475 * 2. Updates the raid superblock when problems encounter.
1476 * 3. Performs writes following reads for array synchronising.
1477 */
1480 {
1487 {
1497 /* drive has already been failed, just ignore any
1498 more fix_read_error() attempts */
1500 }
1508 "md/raid10:%s: %s: Raid device exceeded "
1509 "read_error threshold "
1514 "md/raid10:%s: %s: Failing raid "
1518 }
1519 }
1548 }
1549 sl++;
1556 /* Cannot read from anywhere -- bye bye array */
1560 }
1563 /* write it back and re-read */
1570 sl--;
1583 /* Well, this device is dead */
1585 "md/raid10:%s: read correction "
1586 "write failed"
1597 }
1600 }
1601 }
1607 sl--;
1620 /* Well, this device is dead */
1622 "md/raid10:%s: unable to read back "
1623 "corrected sectors"
1636 "md/raid10:%s: read error corrected"
1642 }
1646 }
1647 }
1652 }
1653 }
1656 {
1676 }
1692 /* we got a read error. Maybe the drive is bad. Maybe just
1693 * the block and we can fix it.
1694 * We freeze all other IO, and try reading the block from
1695 * other devices. When we find one, we re-write
1696 * and check it that fixes the read error.
1697 * This is all done synchronously while the array is
1698 * frozen.
1699 */
1704 }
1738 }
1739 }
1741 }
1744 }
1748 {
1758 }
1760 /*
1761 * perform a "sync" on one "block"
1762 *
1763 * We need to make sure that no normal I/O request - particularly write
1764 * requests - conflict with active sync requests.
1765 *
1766 * This is achieved by tracking pending requests and a 'barrier' concept
1767 * that can be installed to exclude normal IO requests.
1768 *
1769 * Resync and recovery are handled very differently.
1770 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1771 *
1772 * For resync, we iterate over virtual addresses, read all copies,
1773 * and update if there are differences. If only one copy is live,
1774 * skip it.
1775 * For recovery, we iterate over physical addresses, read a good
1776 * value for each non-in_sync drive, and over-write.
1777 *
1778 * So, for recovery we may have several outstanding complex requests for a
1779 * given address, one for each out-of-sync device. We model this by allocating
1780 * a number of r10_bio structures, one for each out-of-sync device.
1781 * As we setup these structures, we collect all bio's together into a list
1782 * which we then process collectively to add pages, and then process again
1783 * to pass to generic_make_request.
1784 *
1785 * The r10_bio structures are linked using a borrowed master_bio pointer.
1786 * This link is counted in ->remaining. When the r10_bio that points to NULL
1787 * has its remaining count decremented to 0, the whole complex operation
1788 * is complete.
1789 *
1790 */
1793 {
1810 skipped:
1815 /* If we aborted, we need to abort the
1816 * sync on the 'current' bitmap chucks (there can
1817 * be several when recovering multiple devices).
1818 * as we may have started syncing it but not finished.
1819 * We can find the current address in
1820 * mddev->curr_resync, but for recovery,
1821 * we need to convert that to several
1822 * virtual addresses.
1823 */
1829 sector_t sect =
1833 }
1841 }
1843 /* if there has been nothing to do on any drive,
1844 * then there is nothing to do at all..
1845 */
1848 }
1853 /* make sure whole request will fit in a chunk - if chunks
1854 * are meaningful
1855 */
1859 /*
1860 * If there is non-resync activity waiting for us then
1861 * put in a delay to throttle resync.
1862 */
1866 /* Again, very different code for resync and recovery.
1867 * Both must result in an r10bio with a list of bios that
1868 * have bi_end_io, bi_sector, bi_bdev set,
1869 * and bi_private set to the r10bio.
1870 * For recovery, we may actually create several r10bios
1871 * with 2 bios in each, that correspond to the bios in the main one.
1872 * In this case, the subordinate r10bios link back through a
1873 * borrowed master_bio pointer, and the counter in the master
1874 * includes a ref from each subordinate.
1875 */
1876 /* First, we decide what to do and set ->bi_end_io
1877 * To end_sync_read if we want to read, and
1878 * end_sync_write if we will want to write.
1879 */
1883 /* recovery... the complicated one */
1891 /* want to reconstruct this device */
1895 /* Unless we are doing a full sync, we only need
1896 * to recover the block if it is set in the bitmap
1897 */
1904 /* yep, skip the sync_blocks here, but don't assume
1905 * that there will never be anything to do here
1906 */
1909 }
1924 /* Need to check if the array will still be
1925 * degraded
1926 */
1932 }
1941 /* This is where we read from */
1953 /* and we write to 'i' */
1973 }
1974 }
1976 /* Cannot recover, so abort the recovery */
1987 }
1988 }
1995 }
1997 }
1999 /* resync. Schedule a read for every block at this virt offset */
2007 /* We can skip this block */
2010 }
2044 count++;
2045 }
2052 }
2056 }
2057 }
2068 }
2084 /* stop here */
2088 /* remove last page from this bio */
2092 }
2094 }
2096 }
2100 bio_full:
2114 }
2115 }
2118 /* pretend they weren't skipped, it makes
2119 * no important difference in this case
2120 */
2124 giveup:
2125 /* There is nowhere to write, so all non-sync
2126 * drives must be failed, so try the next chunk...
2127 */
2132 chunks_skipped ++;
2135 }
2137 static sector_t
2139 {
2140 sector_t size;
2154 }
2158 {
2170 }
2181 }
2189 GFP_KERNEL);
2215 /* 'size' is now the number of chunks in the array */
2216 /* calculate "used chunks per device" in 'stride' */
2219 /* We need to round up when dividing by raid_disks to
2220 * get the stride size.
2221 */
2229 else
2248 out:
2257 }
2259 }
2262 {
2267 sector_t size;
2269 /*
2270 * copy the already verified devices into our private RAID10
2271 * bookkeeping area. [whatever we allocate in run(),
2272 * should be freed in stop()]
2273 */
2280 }
2294 else
2306 /* MOVE 'rd%d' link !! */
2307 }
2313 /* as we don't honour merge_bvec_fn, we must never risk
2314 * violating it, so limit max_segments to 1 lying
2315 * within a single page.
2316 */
2321 }
2324 }
2325 /* need to check that every block has at least one working mirror */
2330 }
2343 }
2344 }
2354 /*
2355 * Ok, everything is just fine now
2356 */
2366 /* Calculate max read-ahead size.
2367 * We need to readahead at least twice a whole stripe....
2368 * maybe...
2369 */
2370 {
2376 }
2383 out_free_conf:
2391 out:
2393 }
2396 {
2411 }
2414 {
2424 }
2425 }
2428 {
2436 }
2438 /* Update slot numbers to obtain
2439 * degraded raid10 with missing mirrors
2440 */
2443 }
2445 /* Set new parameters */
2447 /* new layout: far_copies = 1, near_copies = 2 */
2453 /* make sure it will be not marked as dirty */
2459 }
2462 {
2465 /* raid10 can take over:
2466 * raid0 - providing it has only two drives
2467 */
2469 /* for raid0 takeover only one zone is supported */
2476 }
2478 }
2480 }
2483 {
2499 };
2502 {
2504 }
2507 {
2509 }